AI for Eyes

Ophthalmology is dominated by imaging. Volumetric, three dimensional (3D) ophthalmic imaging using optical coherence tomography (OCT) has revolutionized assessment of the eye and artificial intelligence (AI) improved clinical decision-making.  Current commercial OCT instruments, especially spectral domain (SD) OCT, are widely used in diagnosis and management of patients with retinal diseases.  Yet, standard 2D cross-sectional images of the retina, that remain the most commonly used OCT images and can be even taken by patients themselves, using smartphone apps, can also provide valuable information utilizing AI models. 

Fundus photography - serial photographs of the interior of the eye (opposite the lens)

taken through the pupil by low-power microscope can help to examine optic disc, retina, and lens. With the drastic improvement in smartphone optics, smartphone fundoscopy has been used with increasing frequency since 2010. Machine learning, particularly deep learning, could analyze millions of such images, to identify and quantify pathological features in almost every ophthalmic disease. Even more, it can detect other health conditions such as hypertension, stroke risk, heart disease, and diabetes. 

            

Using deep learning models, systolic blood pressure could be detected as hypertensive with 60% accuracy (Dai t al., 2020) or within 11 mmHg, major cardiac adverse events with accuracy 70% (Poplin et al., 2018) and glaucoma predicted with 96% accuracy (Gheisari et al, 2021). 

The ImageNet dataset - a very large collection of human annotated photographs (over 14 mln)  - is a good starting point for obtaining a model that performs well in recognizing retinal images. A well-known class of deep neural networks - such as a successful CNN trained on ImageNet can be applied to a retinal dataset, and another classifier learns to work with CNN-encoded features - the method known as transfer learning.  Deep learning can be also combined with traditional machine learning methods and fine-tuning approaches. However, many retinal health variables, such as intraocular pressure, cannot be yet adequately predicted from clinical parameters or retinal photographs even using state-of-art molecular learning or deep learning techniques (Ishii et al., 2021). Only one out of three AI-based algorithms designed to detect diabetic retinopathy was able to outperform human screeners. Possibly, we just need more data. But we might be also needing new models.  

Three principal applications of AI for image analysis are classification, segmentation and prediction.  Automated image segmentation and classification can be done without AI methods, just by  applying a set of mathematical functions on the content of an image and classic ML approaches like SVM or random forest. Deep learning approaches could enhance these tasks. One of newer deep learning techniques, generative adversarial network (GAN), can greatly improve resolution of images (super resolution (SR) estimation from a low-resolution counterpart) and image segmentation. GANs can be also used to synthesize images with various eye pathologies, increasing accuracy of classification tasks. 

Thanks to the advances in AI and smart portable or home devices, the future of medicine, including teleophthalmology, is truly exciting. 

REFERENCES

Schmidt-Erfurth U, Sadeghipour A, Gerendas BS, Waldstein SM, Bogunović H. Artificial intelligence in retina. Progress in retinal and eye research. 2018 Nov 1;67:1-29.  

Gheisari S, Shariflou S, Phu J, Kennedy PJ, Agar A, Kalloniatis M, Golzan SM. A combined convolutional and recurrent neural network for enhanced glaucoma detection. Scientific reports. 2021 Jan 21;11(1):1-1.

Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng L, Webster DR. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomedical Engineering. 2018 Mar;2(3):158-64.

Dai G, He W, Xu L, Pazo EE, Lin T, Liu S, Zhang C. Exploring the effect of hypertension on retinal microvasculature using deep learning on East Asian population. PloS one. 2020 Mar 5;15(3):e0230111.

Ishii K, Asaoka R, Omoto T, Mitaki S, Fujino Y, Murata H, Onoda K, Nagai A, Yamaguchi S, Obana A, Tanito M. Predicting intraocular pressure using systemic variables or fundus photography with deep learning in a health examination cohort. Scientific Reports. 2021 Feb 11;11(1):1-0.

Genes and microbes

The body’s assortment of microorganisms depends on what we eat, drugs we take, the stress we are subjected to and the environment we interact with (eg, the infamous SARS-CoV-2 virus). Yet genes also have their say.

In a study of 977 twins in UK, the most heritable taxonomic group of bacteria was found to be Christensenellaceae (Goodrich et al, 2014). These bacteria is present in higher amounts in genetically-lean individuals. It encourages growth of other microbes connected to body weight and energy conservation such as methanogenic Archaea.

A study of over 1500 healthy individuals in Canada (Turpin et al, 2016), associated another abundant bacteria Faecalibacterium with immune system gene CNTN6 (rs1394174), and linked several other genes and bacteria of minor clinical importance (rs59846192 of DMRTB1 - Lachnospira, rs28473221 of SALL3 - Eubacterium, ), rs62171178 nearest UBR3 - Rikenellaceae).

A new paper posted this month on BioRxiv, reports results of a larger genome-wide association study performed for 7,738 individuals from the northern Netherlands.

The authors investigated 5.5 million common genetic variants using linear mixed models on hundreds of bacterial groups and pathways. Potential confounders such as medication usage, anthropometric data and stool characteristics were carefully considered along with dietary information. 

The strongest associations were identified in intronic regions of genes. In particular, between rs182549 in intronic region of the MCM6 gene and Bifidobacteria. This SNP was found to be responsible for lactose intolerance in European population. And so was Bifidobacteria - the most researched and most effective probiotic against lactose intolerance.

Another interesting microbe Collinsella and its family Coriobacteriaceae, associated with rheumatoid arthritis, cholesterol metabolism and leaky gut, was linked to several SNPs regulating genes responsible for the blood group antigens. Blood types does matter. 

Genetic factors might influence our preferences of vegetables, fruit, starchy foods, meat, fish, dairy and snacks. The Dutch study confirmed an earlier finding that rs642387, a genetic variation near genes influencing brain function, is linked to microbial family Rikenellaceae. The paper found that these bacteria, when present in large numbers in the gut, led to decreased consumption of salt. They also showed that an increase in the bacterial pathway of histidine degradation led to increased intake of processed meat.

The paper provides a wealth of information and comes with a lot of supplementary material.

REEFERENES

Goodrich JK, Waters JL, Poole AC, Sutter JL, Koren O, Blekhman R, Beaumont M, Van Treuren W, Knight R, Bell JT, Spector TD. Human genetics shape the gut microbiome. Cell. 2014 Nov 6;159(4):789-99. 

Turpin W, Espin-Garcia O, Xu W, Silverberg MS, Kevans D, Smith MI, Guttman DS, Griffiths A, Panaccione R, Otley A, Xu L. Association of host genome with intestinal microbial composition in a large healthy cohort. Nature genetics. 2016 Nov;48(11):1413.

Lopera-Maya EA, Kurilshikov A, van der Graaf A, Hu S, Andreu-Sánchez S, Chen L, Vila AV, Gacesa R, Sinha T, Collij V, Klaassen MA. Effect of host genetics on the gut microbiome in 7,738 participants of the Dutch Microbiome Project. bioRxiv. 2020 Jan 1.

Sorry I did not quite get that, try again

The holy grail of AI is to fully understand human language in all its nuances. To do that, it should be able to assess, extract and evaluate information from textual data. Were are we now in 2020?